Skip to content

Status.md

Latest commit

 

History

History
1175 lines (1075 loc) · 84.6 KB

Status.md

File metadata and controls

1175 lines (1075 loc) · 84.6 KB
Raw

Project Status and Next Steps

Snapshot: 2026-11-09, after the two Tier C gameplay-depth increments and the 1.6.0 compatibility-completion release; the TEST-07 audio-gating, scripted-input, and nightly baseline update; the 1.5.0 GPU-HQx, video-clean VDLP, field-boundary, and CD-parity update; the 1.4.0 cover-art, latency, unsigned-disc, exact-clock, and differential-harness update; the TEST-07 oracle increment, the VID-12 latency update, the PERF-02 AOT/trimming audit, the 1.3.0 field-handling update, the 1.2.2 performance/library update, the 1.2.1 arm64 packaging update, the 1.2.0 feature/accuracy update, the rewind-capture stall fix, SYS-12 SWI HLE, SYS-17 game-database expansion, SYS-08 clock validation, the third PERF-08 JIT pass, the ENG-02 interpreter hot-path increment, and the PERF-05 high-resolution VDLP copy: Direct3D 11, startup release checks, the fuzz harness, per-edge overscan/void border, screenshot gallery filters, FIQ/DMA/DSP telemetry, flat CHD/raw ISO support, and display-refresh control. The earlier first full-collection sweep, performance, timing-option, BIOS, NVRAM, rewind, cheat, TAS, disc-browser, input, OSD/quick-menu, library, CD-DA, multi-disc, portable, PAL/NTSC, diagnostics, save-state, overscan, and fix-parity passes are retained below. Road Rash's 100k-state race now measures 4.20 ms/frame total (p99 7.40 ms), Debug core builds are optimized, and Auto presentation frame skip is the default. 594 discs tested: 593 Loadable/Rendered, 1 Loadable/No video (known boot failure), 0 Not Loadable.

Corpus update (2026-11-08): the current collection is 1,371 images → 414 unique titles / 378 game test set. Under the compatibility plan's binary beyond-title rule, 375 of 378 games are Playable — the full 147-title peer gap is verified, the comparison has 0 Behind and 0 Regression rows, and 5DO is a strict superset of the 252-title peer union (after the evidenced Return Fire – Maps O' Death data-disc invalidation). The original ≥254 numeric target is exceeded. The two Title-only rows and one No video row are non-retail test images (Joypad Test v1.0, Return Fire – Maps O' Death data disc, and LameNES Emulator). Per-title BIOS, disc hash, capture frame, screenshot, and failure reason live in BeyondTitleEvidence.md and docs/data/beyond-title-results.csv. The unique-title 5DO/4DO/Phoenix comparison is in CompatibilityComparison.md; the 594-disc legacy sweep above remains the per-image baseline.

1.6.0 release update (2026-11-08): v1.6.0 completes the compatibility plan on the current corpus: 1,371 images / 414 unique titles / 378 game-test-set titles, with 375 Playable, 0 Behind/Regression, and a strict superset of the 252-title peer union. The release publishes the authoritative 5DO/4DO/Phoenix matrix and the beyond-title ledger, tightens CD poll reset, guest drive speed, SESSION INFO, TOC Q-position, and volume-header readable-disc behavior, and moves the CD-ROM save-state section to version 5. Settings and NVRAM persistence now fall back to a direct copy when an atomic replace is denied, preserve newer-schema settings as settings.json.v<version>.bak, and survive null/short Players arrays; the desktop adds tooltip coverage and logs expected fallback paths. The suite is 2,061 tests; build, format, and InspectCode are clean; verify-beyond-title-ledger.py and generate-compat-comparison.py --check pass; Road Rash boots for 600 frames in both the compatible and accuracy profiles. The six framework-dependent bundles and SHA256SUMS.txt are published at https://github.com/purelogiccode/5DO/releases/tag/v1.6.0. Release run https://github.com/purelogiccode/5DO/actions/runs/42552318341 passed all six bundle jobs and the publish job; CI run https://github.com/purelogiccode/5DO/actions/runs/42552303788 passed every required job; docs run https://github.com/purelogiccode/5DO/actions/runs/42552303805 deployed Pages and the wiki. All published checksums verify. See WhatsNew.md and ReleaseNotes-1.6.0.md.

Tier C gameplay-depth update (2026-11-09): the first Tier C increment re-ran the 25 titles still graded Playable after Tier B with title-specific start and gameplay scripts. Eight were promoted to Good: Olympic Soccer (full match to FULL-TIME), John Madden Football (live football gameplay through 36,000 frames), Crime Patrol and Mad Dog II (live-action lightgun scenes at 48,000), Shootout at Old Tucson (lightgun scene with the ammo HUD), Guardian War (world-map gameplay through 36,000 frames), Hell - A Cyberpunk Thriller (3D adventure gameplay at 44,000/48,000 frames), and Slayer (first-person dungeon gameplay at 80,000/90,000 frames). A fixed-script run withdrew the earlier PGA Tour 96 promotion; the Player List stays put, so it remains Playable. The CSV ledger moved to 101 rows: 84 Good, 17 Playable. The 378-title playable count stays 375 because Tier C changes the quality grade, not the binary compatibility result. The 25-title baseline completed 25/25. See TestCorpus.md 1.1bm and CompatibilitySweep.md 4.3.

Tier C second increment (2026-11-09): input scripts now support pointer devices (@mouse, @lightgun, @arcadegun, @trackball) with one-frame relative motion and persistent buttons; ScriptedPbusSource applies them per configured port and the parser has six new tests. Four more titles were promoted to Good: Carrier - Fortress at Sea (trap menu to a live trap approach and WAVE OFF at 14,000 frames), Family Feud (home questions and letter keyboard at 9,000/12,000), Insect War (first-person HUD and WARNING state at 58,000/62,000), and N. O. B. (VITALITY HUD gameplay at 9,000). The ledger is now 101 rows: 88 Good, 13 Playable; seven of the remaining rows are media/reference/add-on discs, and six are game leads (BladeForce, DeathKeep, Jurassic Park, Kingdom, Off-world Interceptor, PGA Tour 96). Differential triage closed the three suspected core defects as content/script limits: DeathKeep and Kingdom show the same black screen/title screen on OperaCore, and BladeForce reaches the same Main Options menu on both engines. The quality ledger gate (tools/verify-playability-grades.py and compat-baseline.yml) now fails on a Good -> Playable drop. The suite is 2,068 tests; build, format, and InspectCode are clean. See TestCorpus.md 1.1bo and CliReference.md.

1.5.0 release update (2026-11-08): v1.5.0 moves HQx 2x/3x/4x to shared rule-table GPU passes in OpenGL/GLES, Direct3D 11, and Vulkan, with the CPU filter kept as the reference oracle. The 3DO background is now 24-bit end to end, a zero VDL display-list word is a stop, RunFrame ends on the reference field boundary, and CD poll/FIQ/READ_DATA behavior is closer to Opera. The 50-title panel is 50/50 frame hashes at 60/120/300 with 0 errors: the same-frame instruction delta is 11-20 at frame 60 and 23-2,223 at frame 120, and the remaining behavioral divergence is CD command cadence. The command-50 boot handoff fell from +207,707 cycles / +86,511 instructions to -818 / -376. The suite is 2,033 tests; build, format, and InspectCode are clean. Release run https://github.com/purelogiccode/5DO/actions/runs/42445302517 passed all six bundle jobs and the publish job; CI run https://github.com/purelogiccode/5DO/actions/runs/42445297515 passed every required job. The six bundles and SHA256SUMS.txt are attached at https://github.com/purelogiccode/5DO/releases/tag/v1.5.0, all checksums verify, and local copies are kept in src/FiveDo.Frontend.Avalonia/bin/Release. See WhatsNew.md, ReleaseNotes-1.5.0.md, and DifferentialHarness.md.

TEST-07 audio and scripted-input update (2026-11-08): the differential harness now gates the cumulative audio hash when both engines have produced the same sample count. The 50-title panel matches 150/150 audio fields at frames 60/120/300, so all 50 titles match video, samples, and audio at the panel checkpoints. The one measured long-run limit is the reference clock: 5DO follows the exact 60000/1001 NTSC rate (59.940059940 Hz), while OperaCore's 16.16 constant is 3,928,227/65,536 (59.939987183 Hz), 1.214 ppm lower. On Lemmings the sample counts match through 1,000 frames, then the reference is +1 at 2,000 and +2 at 3,000 frames (2.68 predicted); the harness labels that as an alignment state instead of a waveform divergence. A difference at an aligned count is still a hard failure. Scripted input is wired into both engines: the five-title, 900-frame samples5-input panel matches 5/5 frame hashes at 60/300/600, the generic menu script changes guest output at 600/900, and the no-input control confirms the matches are not from an ignored script. The nightly workflow now runs the showcase fresh, uploads report.html, and fails only when a 5DO field changes against the checked-in baseline. The suite is 2,048 tests; build, format, and InspectCode are clean. See DifferentialHarness.md §4.1/§4.2/§6 and TestCorpus.md 1.1bd.

CORE-04 CD register and metadata parity update (2026-11-08): the debug reference core now records every CD command byte, poll write/read, status read, and data read with cycle and PC, and the harness writes matched fivedo.cdreg.txt/opera.cdreg.txt traces with --cd-register-trace. On 20th Century Video Almanac the two engines' command packets, status bytes, and 2048-byte sector data are identical through command 105, the last command before the sequence divergence; the remaining difference is guest sleep time, not the CD register contract. The trace exposed and fixed the XBUS 2x status bit (MEI_CDROM_DOUBLE_SPEED is 0x80, and the drive starts at double speed), the guest speed cap in cdrom_effective_speed, the readable-disc limit (3DO volume header volume_block_count, not the physical image size), and the TOC Q-position bit plus the 8-byte READ_SESSION_INFO payload. The confirmation panel (test110-panel50-metadata, 50 titles, 300 frames) captures 50/50 with 0 errors, 0 frame-hash divergences, 150/150 aligned audio fields, 39 phase-only titles, and the same 11 CD-only first behavioral divergences at f=300. The CD-ROM save-state section is version 5. The suite is 2,048 tests; build and format are clean. See TestCorpus.md 1.1bf and PortingNotes.md 12.19.

1.4.0 release update (2026-11-07): v1.4.0 adds online cover art (473 unique WebP covers, 645/653 game IDs mapped), VID-12 run-ahead/frame-delay/VRR/audio-sync controls, SYS-06 unsigned-disc support, the exact 60000/1001 NTSC field clock and rational VDL accumulator, VDLP blanking/background fixes, and the live OpenAL device-switching and test-tone fixes. The suite is 1,905 tests; build, format, and InspectCode are clean. The 50-title differential panel remains the current accuracy evidence: 32/50 titles match every frame hash at 60/120/300, every sample delta is within 9, and Space Ace, Wolfenstein 3D, Alone in the Dark, and Lemmings match Opera at 3,000 frames while Road Rash still diverges. See WhatsNew.md, ReleaseNotes-1.4.0.md, DifferentialHarness.md, and CoverArt.md.

Post-1.4.0 differential update (2026-11-07): the 50-title panel is now video-clean: 50/50 titles match every frame hash at frames 60/120/300 (up from 32/50), with 0 errors in test55-panel50-first-sector. 48/50 first PC divergences instruction-align to Opera frame 59 and 2/50 to frame 119 (the f=60 checkpoint itself matched on those two), at a residual of only -55..-415 instructions (same-frame delta ~89,000). The fixes are a 24-bit VDL background (the old 5-bit narrowing rendered near-black as pure black) and treating a zero VDL display-list word as a stop that retains the previous bitmap, matching opera_vdlp.c. The only remaining behavioral divergence is CD command cadence: 36 titles first differ by 1-2 commands at frame 120 (in-flight at the checkpoint) and 13 at frame 300 (seven FMV titles at +3/+4, six at -1/-2). The 5-title panel matches 5/5 video at 300 frames; at 3,000 frames Alone, Lemmings, Space Ace, and Wolfenstein match and Road Rash still diverges. The harness now records pcPhaseFrame/pcPhaseInstructionDelta and per-command PC/retired-instruction stamps. Those stamps showed the earlier command-50 handoff was +207,707 cycles / +86,511 instructions of extra guest work; matching the oracle's first-sector READ_DATA status publication reduced it to -818 / -376, and the f=300 command-104 cycle delta fell from +324,027 to +111,990. READ_DATA deadlines now accumulate from the previous deadline instead of re-basing to now. The residual universal frame-60 one-field phase is separate from that handoff and remains open. The harness also stages ISOs atomically, retries a transient CHD read once, and reports a per-title capture failure as an Error instead of aborting the run. The suite is 1,922 tests; build, format, and InspectCode are clean. See DifferentialHarness.md and TestCorpus.md 1.1az.

CD cadence triage (2026-11-07): the harness now classifies every final CD transcript and reports the largest cumulative cycle-delta step with its command-issue PC and retired-instruction jump (cdCadence, plus a CD cadence column in the reports). The 50-title panel resolves to 39 identical command streams, 8 prefix-only, and 3 one-command READ_ID poll shifts (Chiki Chiki, Dragon's Lair, Virtual Horror). 24/50 titles have a significant step and all 24 issue the step command from the same PC on both engines (17 at 60-100k cycles, six at 20-60k, one >=100k). The fresh five-title run confirms Alone 361/361 (+379 cycles at the last command), Road Rash 395/395 (+468), Wolfenstein 341/341 (-345), Lemmings 109/105 with one +22,834 step at command 51 (+8,242 instructions), and Space Ace 168/168 with one +80,567 step at command 164 (+32,330 instructions). The residual count difference is guest wait-loop cadence, not a divergent CD command table. The new --cd-register-trace shows no CD register access during the 237,794-cycle Space Ace wait or the 2.17M-cycle Lemmings wait, so the guest is waiting on a non-CD source (timer/VBlank); this moves the next measurement to CORE-07. The suite is 1,922 tests. See DifferentialHarness.md 8 and TestCorpus.md 1.1bb.

Frame-boundary parity (2026-11-07): the universal frame-60 PC/CPSR phase had two causes. The harness finalized each Opera checkpoint at the next video callback, so ARM/CD state belonged to the following field while the audio hash was already correct; the harness now captures ARM/CD state in the checkpoint's own callback. FiveDoMachine.RunFrame also stopped on the rounded 208,541-cycle budget while Opera stops on the 263-line scanline boundary, so the discarded last-batch overshoot accumulated about 3.3 instructions per field; RunFrame now ends on Quartz.VdlFieldIndex. On the 50-title confirmation panel the same-frame retired-instruction delta is 11-20 at f=60 (average 15.3) and 23-2,223 at f=120, with 50/50 frame hashes and 0 errors. The CD count matches on 50/50 titles at f=60 and f=120 (down from 36 first divergences at f=120), and the first behavioral divergence moved to 11 titles at f=300 (8 count, 3 sequence); the other 39 titles are phase-only. This supersedes the "5DO frame 60 aligns with Opera frame 59" interpretation. The 3,000-frame showcase keeps 4/5 video matches (Road Rash is the only frame-hash divergence) with CD counts Lemmings 365/352, Gex 548/547, Road Rash 851/836, Wolfenstein 862/861, and Alone 932/932. The suite is 1,923 tests. See DifferentialHarness.md 8 and TestCorpus.md 1.1bc.

Release workflow https://github.com/purelogiccode/5DO/actions/runs/42399055155 passed all six bundle jobs and published six versioned archives plus SHA256SUMS.txt; all six checksums verify. CI run https://github.com/purelogiccode/5DO/actions/runs/42398646569 passed every required job, including the six bundle builds, NativeAOT CLI/frontend, format, tests, and benchmark gates. The release is published at https://github.com/purelogiccode/5DO/releases/tag/v1.4.0. Local bundles are kept under src/FiveDo.Frontend.Avalonia/bin/Release.

TEST-07 exact-clock and 50-title panel update (2026-11-06): the earlier claim that the frame-60 CPSR difference was the FIQ-disable bit was wrong: 0x600000d3 and 0x800000d3 differ in the N/Z/C condition flags. The real defect was the field clock. 5DO now runs NTSC at the exact 60000/1001 field rate (208,541 cycles at 12.5 MHz) and uses a rational VDL accumulator, removing both the 1 ms/60-field drift and the per-batch 8.24 rounding that ran the scanline clock ~0.09% fast. Sample counts now track Opera within 1-8 samples at 300 frames (all 50 panel titles are within 9 at frame 300), and 32/50 titles match every frame hash at frames 60/120/300. The 50-title panel captured 50/50 with 0 errors; CD counts match 49/50 at frame 60. The absolute instruction count carries a fixed ~89,000 startup offset, but the per-frame retirement rate matches within 0.1% on 4/5 titles, so PC/CPSR is a phase sample rather than a cycle-model gap. The new CD cycle stamps localize the remaining CD cadence difference. At 3,000 frames Alone, Lemmings, Space Ace, and Wolfenstein match Opera's hash; Road Rash still diverges. The suite is now 1,905 tests. See DifferentialHarness.md and TestCorpus.md 1.1ax. TEST-07 oracle increment (2026-11-06, superseded by the exact-clock update above; the FIQ-bit and 59.94/60 Hz statements below are incorrect): the differential harness now applies a fixed seed to both engines (--seed, default 1488), compares ARM CPSR and CD command counts/opcode hashes, and reports the first divergence in the console, JSON, and Markdown. Two consecutive runs of the new five-title benchmarks/diff/samples5.json panel produced identical checkpoint arrays for both engines. The first divergence is the ARM PC at frame 60 on all five titles; Lemmings' CD commands match 9/9 at frame 60 and 50/50 at frame 120, then 108/105 at frame 300, and its CPSR differs at frame 60 by the FIQ-disable bit. The cumulative audio hash is recorded but not gated because the 59.94/60 Hz field-rate offset leaves a systematic sample-count delta. The panel command generated benchmarks/diff/panel50.json (50 evenly spaced titles from the 594-disc manifest). A 3,000-frame follow-up separates PC drift from visible divergence: Lemmings matches the frame hash at every checkpoint, while Road Rash and Space Ace show different animation frames/scenes and no ±60-frame hash match. The full CD command sequence localizes the first command divergence to READ_DATA (0x10) versus READ_ID (0x83) transitions, giving CORE-04 a concrete next target. The suite is 1,881 tests. See DifferentialHarness.md 8 and TestCorpus.md 1.1aw.

Performance increment (2026-11-05): the default rewind history no longer stalls the emulation thread. Keyframe capture on Lemmings fell from 139.3 ms to 7.4-9.2 ms (one 14.1 ms outlier), per-1,200-frame allocations fell from ~1.5 GB to 16 MB with no gen0/gen2 collections, and the frame loop measured p50 1.34 ms / p99 6.85 ms / max 14.76 ms with the mailbox and rewind active. The memory save-state section (version 4) now stores only the live RAM/VRAM range, cutting a stock-layout uncompressed state from ~26 MB to ~10 MB, and the frontend opens discs through the prefetch cache. See Performance.md 8.5. The suite is 1,866 tests.

CPU and high-res increment (2026-11-06): the third PERF-08 pass gives compiled blocks a shared PC/CPSR/cycle/instruction context, compiles register-specified shifts, and replaces the compile-hit dictionary with a direct-mapped table. Lemmings improved from 1.60-1.66 to 1.23-1.37 ms/frame (about 20% faster than the old JIT) but is still ~1.25x slower than the 1.00-1.02 ms/frame interpreter, so the interpreter remains the default; the synthetic loop is 0.89-0.99 ms/frame versus 1.20-1.30 for the interpreter. The second pass had replaced per-instruction JIT verification with exact RAM-write block invalidation and fixed compiled R15 operands. ENG-02 added direct RAM/ROM instruction fetch, inline LSL#0/rotated-immediate operand paths, a flag-preserving ALU dispatch, and population-count LDM/STM iteration; an interleaved A/B measured synthetic 1.58→1.33 ms/frame and Lemmings 1.43→1.13 ms/frame with unchanged hashes. PERF-05 now reads the raw VRAM word span for high-resolution scanline DMA, cutting Lemmings high-res from 2.42-2.47 to 1.30-1.43 ms/frame (about 1.8x) with an identical high-res hash. See Performance.md 8.4 and 8.6.

Latency and AOT increment (2026-11-06): VID-12 adds 0-2 frame run-ahead, 0-8 frame presentation delay, VRR-friendly compositor presentation, a −250…250 ms audio sync offset, per-game profile fields, and software input-latency telemetry. Run-ahead uses a reusable uncompressed MachineStateSnapshot; bench --latency measures a 9,348,438-byte snapshot, ~3.6-3.9 ms capture, and ~1.6-2.0 ms restore. AOT CLI/frontend win-x64 builds start, render, play audio, and pass smoke tests; the AOT frontend booted Lemmings with run-ahead 1, frame delay 1, VRR, and audio sync 40 ms. PERF-02 is complete; VID-12 and SYS-12 are marked done. Full optical input-to-photon measurement and real adaptive-sync hardware confirmation remain documented validation caveats. See Performance.md 8.7 and Trimming.md.

Release status (2026-11-02): v1.0.0 is released as a documented-compatibility release. The version is bumped, the source is tagged v1.0.0, and one unified bundle per system (Windows, Linux, macOS x64) plus checksums is attached to the GitHub release. Linux x64 validation passed on Linux Mint 22.3: the full test suite (1,473/1,473), selftest, bench, a 3,000-frame Wolfenstein boot, non-silent audio, and the frontend --smoke-test. The GL probe timed out under the VM's llvmpipe/SSH session, so Linux GPU presentation is a preview/best-effort limitation. Legal/attribution sign-off is recorded under the FreeDO modified GNU LGPL as free, open-source donationware (non-commercial; voluntary donations only). M13 loadability is complete; M13.5/M13.6 visual/high-resolution review and M14 high-resolution CEL research are post-1.0.

Field-handling update (2026-11-06): v1.3.0 was the previous release, published at https://github.com/purelogiccode/5DO/releases/tag/v1.3.0. It adds InterlaceProcessor with auto 480i detection, bob, weave, and blend modes; per-game and --deinterlace selection; interlace-aware viewport geometry; screenshot/GIF capture of the presented frame; and run --json mode/detection fields. The suite is 1,866 tests. Bob-vs-auto hashes match on five local titles (3,000 frames for Road Rash, Space Ace, and Alone in the Dark; 300 for Wolfenstein 3D and Lemmings), forced weave produces a 320x480 frame, and high-resolution output remains 640x480. The CLI and core default to bob; the desktop defaults to auto. No local title is confirmed 480i, so synthetic complementary field tests are the current 480i evidence.

Release workflow https://github.com/purelogiccode/5DO/actions/runs/42235476257 passed all six bundle jobs and published six versioned archives plus SHA256SUMS.txt; the docs workflow https://github.com/purelogiccode/5DO/actions/runs/42235456021 deployed the Pages site and wiki. All release checksums verify. Local bundles and an archive of every historical release asset are kept under src/FiveDo.Frontend.Avalonia/bin/Release.

Performance and library update (2026-11-06): v1.2.2 was the previous release, published at https://github.com/purelogiccode/5DO/releases/tag/v1.2.2. It adds the interpreter, high-resolution VDLP, and rewind performance work; the opt-in managed JIT; the expanded game database and library health/DAT verification; eight Opera memory layouts; eight PBUS ports; opt-in Portfolio OS SWI HLE; and audio-only CHD cataloging. The suite is 1,815 tests. Release workflow https://github.com/purelogiccode/5DO/actions/runs/42169040644 passed all six bundle jobs and published six versioned archives plus SHA256SUMS.txt; all six checksums verify. The docs workflow deployed the Pages site and published the wiki. A post-1.2.1 audit fixed eight bugs: the JIT register-shift PC delta, stale ROM-bank and soft-reset blocks, FIQ entry telemetry, the prefetch track-sector length, legacy PBUS slots 7-8, the pointer-exit port bound, and the ARM clock range.

Arm64 packaging update (2026-11-03): v1.2.1 was the arm64 packaging release, published at https://github.com/purelogiccode/5DO/releases/tag/v1.2.1. It added framework-dependent Windows arm64, Linux arm64, and macOS arm64 bundles alongside the x64 bundles, for six versioned archives plus SHA256SUMS.txt. CI verifies each arm64 apphost's PE/ELF/Mach-O machine type; the x64 bundles keep their runtime smoke tests. No emulator behavior changed. Release workflow https://github.com/purelogiccode/5DO/actions/runs/41583317988 passed all six bundle jobs; every checksum was verified after publication.

Feature update (2026-11-03): v1.2.0 added the optional Windows Direct3D 11 presenter, startup GitHub release checks, deterministic fuzzing, per-edge overscan/void border, screenshot gallery filters/compare/export, display-refresh control, flat 2048-byte CHD and raw .iso detection, FIQ request-to-vector latency telemetry, DSP per-tick statistics, and the CLIO/DIPIR/FIFO/MCTL/XBUS audit fixes. The 1.2.0 suite was 1,597/1,597 passing. WebAssembly/browser, Android/iOS, and Raspberry Pi were removed from the roadmap; desktop Windows/Linux/macOS remain supported. The three versioned bundles and SHA256SUMS.txt are attached to the release; every checksum was verified after publication.

Maintenance update (2026-11-02): v1.1.0 remains available at https://github.com/purelogiccode/5DO/releases/tag/v1.1.0. It adds game library deduplication by 3DO game ID, moves the OpenAL license text to docs/OpenAL-Soft-COPYING.txt, removes the frame/PC/cycle counter from the desktop status bar, and completes the Serilog catch-block coverage and launch-statistics wiring. The 1.1.0 suite was 1,597/1,597 passing. The bundle layout is unchanged: framework-dependent single-file apphosts, no bundled .NET runtime, no PDB/XML files.

Latest work (2026-11-05): Settings > Paths now edits game library folders directly (type or paste a path, Browse, or Remove), and the scanner indexes cover images once per folder instead of probing every cover extension for every disc. Cover art now decodes on bounded background tasks, so the non-virtualizing grid does not block the list. On the local 2,732-image collection the cached listing fell from 2.8 s to 0.33 s, and a cold listing shows all entries in 17 ms before background metadata hashing. The library now reports both counts (for example 653 games · 1,366 images) because discs are identified by the 4DO game ID (MD5 of the first two data sectors); Settings > Paths > Show duplicate disc images lists every renamed or GoodTools/Redump/No-Intro copy. PERF-08 added an opt-in managed IL JIT (--cpu-backend jit) for hot data-processing, load/store, and branch blocks. A third pass shares a PC/CPSR/cycle context across blocks, compiles register-specified shifts, and uses a direct-mapped compile-hit cache. It is bit-identical on the test corpus and a 500-frame CLI A/B matches Road Rash, Powers Kingdom, and Lemmings hashes/instructions/cycles, but Lemmings boot improved only to 1.23-1.37 ms/frame and is still ~1.25x slower than the interpreter, so PERF-08 remains Partial and the interpreter remains the default. SYS-17 expanded the game database from 518 to 653 records and added gamedb stats|scan|export|merge, docs/GameDatabase.md, and a filename parser for Redump/GoodTools/No-Intro layouts. The schema now carries revision, disc number, BIOS/BIOS2, fix mode, input devices, NVRAM, and cheat fields; fix, BIOS, and input values are consumed at runtime. gamedb stats --folder "J:\Panasonic 3DO" reports 1,366 readable images, 653 unique IDs, and 100% metadata coverage with no errors (the three exceptions are audio-only [CD-Audio] CHDs). SYS-08 added the five-title 0.5x-1x-2x-4x A/B in Performance.md §8.3 and kept the stock 12.5 MHz default; every non-stock clock changed guest behavior. The suite is 1,866 tests.

SYS-12 (2026-11-04): Portfolio OS SWI HLE is available behind --swi-hle, Settings > Emulation, and per-game profiles. It handles Opera's 14 fixed-point math SWIs on the host, records the option in save-state Options v9, and reports per-comment coverage with --swi-comments. The host routines match opera_fixedpoint_math.c; a real-BIOS 4x4 benchmark cuts 847,860 guest steps to 24,578. HLE is an opt-in performance mode and skips the guest handler's cycles/register clobbers, so some titles change behavior. Default-off hashes are unchanged. Input devices also support eight PBUS ports: ports 7-8 use the Opera daisy chain, --active-devices/Settings caps the polled ports at 0-8, and the MADAM PBUS buffer matches Opera's 256 bytes (state section v4). The hide-lightgun-crosshair option is also available.

SYS-07 (2026-11-04): all eight Opera DRAM/VRAM layouts (2/4/8/14/15 MB DRAM and 1/2 MB VRAM) are selectable behind --memory, Settings > Emulation > Memory, per-game profiles, CLI JSON, and save states. VRAM is decoded directly above DRAM, VDLP/SPORT/CEL use the live layout, and MADAM RED SYSBITS reports the capacity to the BIOS. A homebrew probe ROM uses 4 MB DRAM + 2 MB VRAM; retail titles remain on the stock 2 MB/1 MB default because Opera documents expanded memory as homebrew-only. The CPU-clock matrix (0.5x-4x) and bench --cpu-clock/--memory were also added; the suite is now 1,866 tests. See TestCorpus.md 1.1al, 1.1an, 1.1ao, 1.1aq, 1.1ar, 1.1as, 1.1at, and Performance.md 8.1.

Latest work (2026-11-02): the CLI and desktop shell check the public GitHub release feed at startup (cached, opt-out) and offer the download page; 5do update exposes the same check. The optional FiveDo.Video.Direct3D Windows backend adds a FL11_0 two-stage presenter, runtime HLSL, a DXGI flip-model child swapchain, --gl-backend d3d11, Settings selection, and --d3d-probe/--d3d-window-probe. 320x240 Accurate matches the CPU oracle (0 different pixels); 4K sharp-bilinear measures present p99 0.23 ms and GPU p99 0.07 ms. OpenGL remains the default.

The 2026-11-02 hardening pass added tools/FiveDo.Fuzz (state, disc, CUE, Opera-FS, CD command-stream, and ARM60 targets plus a nightly workflow), per-edge overscan and per-game crop profiles with a configurable void border, and screenshot gallery filters, side-by-side compare, export, and settings sidecars. The same pass added FIQ request-to-vector latency telemetry and --fiq-trace request/entry lines, closed the CLIO output-FIFO/DIPIR/MADAM-MCTL/XBUS poll audit gaps, and exposed DSP per-tick statistics in run --json. The 1.2.0 suite was 1,597 tests.

SYS-01 and AUD-05 now expose multi-track CUE/CHD tables and CD-DA: ITrackedDiscSource carries native 2048/2336/2352-byte sectors, CueDiscImage handles multi-file/multi-track INDEX/PREGAP/POSTGAP layouts, ChdDiscImage exposes audio tracks, and CdromDevice implements multi-track READ TOC/lead-out, READ SUB-Q, MODE SET/SENSE, PLAY_AUDIO position tracking, 2352-byte digital-audio reads, and guest inject. CD-DA is mixed into the 44.1 kHz DSP output with persisted volume/mute; CD state v4 stores the mode, audio position, volume, and mute. disc tracks reports every track, and SYS-02 adds an M3U/M3U8 playlist, playlist editor, drag-and-drop import, File > Swap Disc/F3, guest-eject ODE auto-advance, and run --playlist/--swap-at/--swap-disc. SYS-03 adds NTSC/PAL1/PAL2 timing: PAL runs at 50 Hz with 288 visible lines, a 312-line field, and region-aware VDLP/machine timing. Battle Chess Europe renders correctly at 320x288. Per-game region overrides and state-driven region switching are supported. CORE-04 now returns illegal-request statuses for unsupported/unknown CD commands, implements the guest RESET command, exposes CdromCommandTranscript/run --cd-transcript, and read-aheads native audio sectors in the disc prefetch cache. FND-02 is complete with settings schema v2, legacy normalization, malformed-file backup, per-game profiles, and settings validate|migrate. UX-15 adds portable mode: portable.flag, --portable, FIVEDO_DATA_DIR, and named profiles redirect settings, NVRAM, cheats, library state, and logs; InstanceLock prevents two desktop processes from writing the same directory. F2 toggles the frame/lag/skip and selected-player input overlay. AUD-01 adds OpenAL device selection and a 20-500 ms buffer target with live output recreation; audio devices lists playback devices. FND-08/UX-17 add diagnostics bundle and Help > Save Diagnostics Bundle... with path/account scrubbing. Save-state format v3 now declares feature flags and a section table, keeps v1/v2 readable, records the live region, and supports warn/reject unknown sections. Remaining disc work is known-title CD-DA evidence and a real in-game multi-disc swap, plus subcode. A 10-disc Release smoke after the TOC change reported 10/10 Loadable/Rendered. 1,473 tests pass. The five not-loadable rows are resolved or explained: Mirai Shounen Conan now runs (transient CHD read error), and the four Arcade CD-ROM titles load with panafz10-patched.bin plus the default headless pad. The extended-boot retry renders all 18 former failures except one: 593 of 594 discs now render, and Crash 'n Burn (Japan) remains a known boot failure. WorldCup Super Stadium needs the FZ-10J BIOS plus its Kanji font ROM (BIOS2), which the BIOS library now classifies correctly. Headless run/trace now attach a default player-1 pad, fixing Olympic Summer Games' controller check. Fifteen triaged titles now have a reviewed 5DO status in GameCompatibilityList.md; the rest remain No data. The CORE-01 ARM timing option (--arm-timing freedo|arm60), its save-state snapshot, --cpu-clock 0.5x-4x, and --random-seed are implemented with tests. run --cycle-classes adds a per-instruction-class instruction/cycle ledger (cycleClasses in run --json), and the exact N=2/N=4 class-cost table is documented in PortingNotes.md 12.28. A 40-title A/B reports 17/40 matches with ARM60 versus 15/40 with FreeDO (TestCorpus.md 1.1a1); the differential harness now accepts --arm-timing, --cel-timing, and --cd-timing alongside --clio-timers and --profile. The CORE-02 CLIO timer option (--clio-timers freedo|opera) now implements Opera's per-slot scan and cascade carry behind the safe FreeDO-compatible default, with a state snapshot, 16-bit timer-register masking in Opera mode, a reference-model conformance test, per-timer telemetry (--clio-timer-stats), a bounded --clio-timer-trace, and the smoke/A-B differentials in TestCorpus.md 1.1b. The reference debug core now records matching timer/read events (--timer-trace), which show the timer-14 period equal at 12,384/12,396 cycles (TestCorpus.md 1.1az). The CORE-03 CEL option (--cel-timing freedo|opera) now implements Opera's deferred draw and pixel/texel busy deadline behind the unchanged synchronous default, with a Madam section v2 busy snapshot, an Options v3 mode snapshot, and a three-title 3,000-frame smoke differential in TestCorpus.md 1.1c. The CORE-04 CD option (--cd-timing freedo|opera) now implements Opera's cycle-based 2048-byte data-sector readiness at the stock 2x drive speed, including delayed data polling, deferred read-data status, the XBUS CD FIQ, a CD-ROM section v2 clock snapshot, and an Options v4 mode snapshot. The immediate FreeDO path remains the default; the three-title smoke differential is in TestCorpus.md 1.1d. The Opera poll reset now keeps the three interrupt enables (0x87), and the XBUS CD FIQ is level-held each clock batch as in opera_xbus_tick; TestCorpus.md 1.1az records the fix and the remaining guest copy/decrypt wait. SYS-09 adds --cd-speed unbounded|1x|2x|4x|8x|16x|24x|32x (Settings > Emulation > CD speed), Opera's per-speed overhead table, an Options v5 speed snapshot with a mismatch warning, and the three-title 3,000-frame smoke differential in TestCorpus.md 1.1e. CORE-08 adds EmulationProfile (Compatible, Accuracy, Performance, Custom), --profile, and Settings > Emulation > Timing profile; the accuracy profile enables the ARM60, Opera timer, CEL busy, and CD cycle options together, and run --json reports the profile derived from the actual options. bench --profile measures any named profile and reports its timing modes, selftest checks determinism plus state round-trip for every named profile, and the differential harness accepts --profile for a full-collection A/B. A 40-title A/B reports 17/40 matches for both compatible and accuracy with identical per-title statuses (TestCorpus.md 1.1ak1); the 594-title A/B remains. SYS-14 adds per-game NVRAM and versions. NvramStorageMode (Global, PerGame), NvramVersion (0-9), and NvramDirectory select shared nvram.ram, versioned nvram.<n>.ram, or nvram/per-game/<gameId>.<n>.ram images. The store reads migration fallbacks from per-game version 0, the shared version, and the legacy image; saves stay atomic, and a corrupt primary is backed up rather than replaced by an older image. Settings > Emulation exposes the mode and version. The CLI adds --nvram-mode, --nvram-version, and --nvram-dir, reports the resolved path in run --json, and adds nvram path|list|export|import|reset. SYS-04 adds the BIOS manager: a built-in MD5 catalog, BiosLibrary scanning/validation, region/Kanji/Anvil/patch metadata, --bios2 and --bios-auto, bios list|scan|select, Settings > Paths BIOS folder/auto-select/scan, per-game BIOS overrides, and the Settings > Emulation override combo. Auto-selection uses the game database region plus known requirements for Slayer Europe/USA, Powers Kingdom (Europe), and Arcade CD-ROM discs; the configured BIOS remains the fallback when no BIOS folder is set. The five-title 3,000-frame smoke run and four known-requirement 12,000-frame runs are in TestCorpus.md 1.1f. The first Opera parity batch (VDLP, ARM carry, DSP FIFO, MADAM span/MARIA/BPP0/SKIPX/LRFORM/blue-LSB/REGCTL/PIXC, CLIO FIFO), OP-14 (XYPOSH/XYPOSL origin, retired GraphicsStepY/BitTiming6), OP-SYS-02 (Opera CD poll/local reset), OP-SYS-01/03's ROM overlay, DIPIR, XBUS, and CSTAT soft-reset work, and OP-SYS-06's separate CLIO PRNG are complete (the DIPIR/soft-reset path remains opt-in pending a boot transcript); see FixParity.md and the completed log in ImplementationPlanArchive.md. UX-01/UX-02 add rewind and fast-forward. RewindBuffer keeps compressed keyframes plus the per-frame PBUS packet, replays input deterministically after the nearest keyframe, clears on timeline jumps, and enforces a 256 MB default cap; CoreRunner.Rewind restores the machine and presents the frame. CoreRunner.SpeedMultiplier runs 2x-10x frames per presentation tick with intermediate rendering skipped and audio muted. The shell binds R (one-second rewind), Shift+F9 (one-frame rewind), and hold-Tab (fast-forward); Settings > Emulation controls the history length and multiplier. UX-04 adds per-game RAM cheats: FiveDo.Cheats parses raw, typed, conditional, and button-activated codes; CheatStore owns the per-game JSON files; FiveDoMachine.BeforeFrame applies them every frame (including rewind replay and fast-forward); the shell adds Emulation > Cheats... and Settings > Paths > Cheats; the CLI adds --cheat, --cheat-code, --cheat-auto, --cheats-dir, and cheat list|add|enable|disable|remove|import|export. UX-06 adds deterministic TAS movies: FiveDo.Tas owns the versioned .5domovie container (per-frame PBUS packets, optional starting state, CRC-32 and Brotli), record/replay, branch/trim, and verification; the CLI adds --movie-record, --movie-play, --movie-verify, and movie info|record|play|verify|branch; the shell adds Emulation > Record Movie.../Play Movie.../Stop Movie and a live player-1 input display. UX-11 adds a disc browser and disc-info dialog: FiveDo.Host.DiscBrowser mounts the Opera file system without starting the machine, shows a sorted directories-first tree with sizes/LBAs, extracts a file, directory, or the whole disc with safe path containment, and reports the volume label/comment, game ID, 4DO database metadata, and automatic fixes. File > Disc Browser... and File > Disc Info... expose it in the shell. SYS-15 adds the missing 3DO input devices: InputDeviceSet selects a joypad, flightstick, mouse, lightgun, arcade lightgun, Orbatak trackball, or no device on each of six ports; PbusChainEncoder emits Opera's daisy-chain records and keeps the legacy 4DO packet while every port is a joypad. The desktop shell adds Settings > Emulation > Input devices, pointer-driven mouse/lightgun input with a crosshair, and the CLI adds --input-device and --pbus-format. Crime Patrol, Mad Dog II, Orbatak, and Shootout at Old Tucson boot and detect the selected arcade device; scripted gameplay and calibration remain. UX-12 adds File > BIOS / NVRAM Manager..., a two-tab manager that scans and classifies BIOS images, imports validated .bin/.rom files, sets BIOS1/BIOS2, pins or clears per-game BIOS assignments, and lists shared/per-game NVRAM images with validation, game metadata, version, size, and modification time. The NVRAM tab backs up one image or the whole layout, imports a validated 32 KB image, resets an image to a fresh volume, and deletes an image with a second-click confirmation. BiosInventory and NvramInventory isolate the file operations for tests and use atomic writes. UX-19 adds OSD notifications and the in-game quick menu. FiveDo.Host.Osd owns a bounded, time-expiring notification list and validated position/opacity/duration options; the game view shows colour-coded toasts for save/load, slot changes, rewind, fast-forward, screenshots, reset, and settings. F4 or the gamepad Select+Start chord opens a quick menu that pauses the game and offers resume/pause, save/load, rewind, fast-forward, screenshot, reset, cheats, settings, stop, and close; Up/Down and Enter navigate, and closing resumes the prior state. Settings > Video exposes the OSD and quick-menu options. UX-13 adds persistent library favorites, launch counts, last-played timestamps, and play time (library-state.json); a 60-second flush bounds crash loss. LibraryPlaytimeTracker measures only running emulation. An optional offline library-metadata.json overlay adds title/developer/publisher/ year/genre/region/players/description by game ID or path, while the 4DO database now supplies year/publisher/region directly. The game list adds star toggles, metadata/play-time text, All/Favorites/Played/Unplayed filters, and Title/Recently played/Most played/Release year/Publisher sorts; Settings > Paths chooses the metadata file. External metadata providers, rich media, and the >=90% rich-metadata target remain. The Lemmings 300-frame differential smoke still passes (1 match, 0 diverged). 1381 tests pass. The differential harness (FND-05/CORE-11/CORE-16) is implemented: tools/FiveDo.DiffHarness stages the bundled OperaCore and BIOS, runs both engines with identical options, compares checkpoint frame hashes, applies manifest allowances, and writes JSON/Markdown/HTML reports. The Lemmings (USA) 300-frame smoke matches at frame 300 (0x110E7DC5 on both engines); the five-title showcase manifest is benchmarks/diff/showcase.json, and its first 3,000-frame run video-matched Wolfenstein 3D and Alone in the Dark while Lemmings, Gex, and Road Rash diverged. The strict PC gate shows 0/5 matches: the video matches still have PC deltas of 184,696 and 739,184 bytes, so frame hashes alone are not sufficient evidence. The manifest command generated the portable 594-title benchmarks/diff/full.json. See DifferentialHarness.md and TestCorpus.md 1.1g. The remaining work is the full-collection run, state/CD transcript fields, and the nightly baseline. The v1.0 documented-compatibility release is ready: M13 loadability is complete (593/594 render), legal/attribution sign-off is recorded, and M13.5/M13.6 plus M14 are post-1.0. The project is donationware (non-commercial, source-available, voluntary donations only). See CompatibilitySweep.md, GameCompatibilityList.md, and HighResolutionCelResearch.md. Release definition: documented compatibility, not full compatibility; see ReleaseChecklist.md. Scope: pure C# 3DO emulator, desktop GL/GLES presenter, optional Vulkan and Direct3D 11 backends, CLI, and deterministic tests. Reference: References/opera-libretro-master/libopera is now the primary behavioral reference for new compatibility work; see ReferenceStrategy.md.

2026-10-30 priority change (superseded 2026-11-02): the v1.0 release was deferred for the M13 full-collection loadability and in-game sweep and M14 high-resolution CEL research. M13 loadability is now complete (594 tested, 593 render, 1 known boot failure, 0 not loadable); M13.5/M13.6 and M14 are post-1.0. Gameplay-depth work (full races, inventory interaction, long playthroughs) remains post-1.0. The project is donationware: non-commercial, source-available, and funded by voluntary donations rather than sales. The Linux validation target is the local Hyper-V VM LinuxMint (D:\Hyper-V\VMs\LinuxMint). See CompatibilitySweep.md, GameCompatibilityList.md, and HighResolutionCelResearch.md.

This is the short, current answer to "where are we?" The documentation index is documentation index. The detailed task list is in ImplementationPlan.md, the performance contract in Performance.md, the local game corpus in TestCorpus.md, the release gate in ReleaseChecklist.md, and the compatibility completion plan/next actions in ImplementationPlan.md.

1. What works today

  • Core: ARM60 interpreter, configurable 2-15 MB DRAM / 1-2 MB VRAM (stock 3 MB total) with 2 MB ROM, shared/per-game NVRAM, CLIO timers/FIQ/FIFO/DMA, MADAM CEL/pixel/matrix engines, VDLP, SPORT, DSP, Quartz scheduler, XBUS, CD-ROM, ISO/BIN/CUE/CHD storage, input/PBUS, game database, deterministic save states.
  • Save states: state format version 2 compresses large section payloads with Brotli while keeping the CRC over the uncompressed data. Version-1 states still load. A 24,902,592-byte local state becomes 647,647 bytes (300-frame Gex) or 2,001,229 bytes (14,900-frame Gex); the section inspector reads the header without decompressing payloads, and state inspect reports stored/compressed sizes.
  • BIOS: panafz1.bin boots, runs the splash/animation, initializes the OS, and reaches the shell. The earlier PC 0x205C4 "halt" was the Portfolio idle loop. SYS-04 adds a BIOS manager: BiosLibrary scans .bin/.rom files, computes MD5, and classifies region, video standard, Kanji, Anvil, and patch state; the built-in catalog covers the common FZ-1/FZ-10/FZ-10E/FZ-10JA, GoldStar, Sanyo, patched arcade, Kanji font, and arcade support images. --bios-auto and Settings > Paths > Auto-select BIOS choose a region-appropriate primary plus optional BIOS2; --bios2 loads an explicit secondary. bios list|scan|select expose the catalog, and per-game BIOS overrides are stored by 4DO game ID. The configured BiosPath remains the fallback, so existing settings keep their exact BIOS.
  • Video: the desktop two-stage presenter expands palettes at native resolution into an FBO, then scales/filters to the output. Desktop GL 3.3 and GLSL ES 300 sources are selected from the context version; the frontend can force the backend with --gl-backend auto|wgl|angle|software. Windows ANGLE/EGL OpenGL ES 3.0 is validated on the development RTX 4070: all filters compile and a 320x240 Accurate capture matches the CPU oracle pixel-for-pixel. The VDLP bitmap path now honors the modulo-derived line width (320/384/512/640/1024). The palette shader was fixed in the eighth pass to use independent red, green, and blue CLUT indices; the previous shader used only the low five bits and made real games grayscale. An opt-in high-resolution mode (--high-res, EmulatorTuning.HighResolution) renders CELs and bitmap lines into four 1 MB VRAM planes and presents 640x480 by line-doubling the 640x240 source; Lemmings' title screen now matches the 640x480 reference. This is not true vertical 640x480 rendering; the internal high-resolution research is M14. The mode is off by default because 320x240 titles must not be upscaled, and the desktop shell exposes it as a Hi-Res checkbox that restarts the machine. An opt-in xBR-lv2 4x filter, HQx 2x/3x/4x filters, and CRT bloom are available in Settings; xBR and HQx are GPU passes over the GPU-expanded native source with sharp-bilinear finishing, and CRT bloom adds a 3x3 native-resolution glow in the CRT filter. The CPU HQx reference hashes remain the oracle, and the GPU HQx rule table matches them in the opt-in parity tests.
  • Performance: the machine runs on a dedicated core thread paced by the 44.1 kHz audio clock (60 Hz fallback) with a frame pacer; the UI renders on demand; uploads use TexStorage2D + TexSubImage2D; uniform locations and the present delegate are cached; physical-pixel sizing is implemented and unit-tested at 100/125/150/200% scaling. The benchmark reports p50/p95/p99 and per-stage allocations, and the synthetic steady state is 0 bytes/frame. An opt-in bounded prefetch cache is available for CHD/ISO streaming. The --gl-probe validates both WGL and ANGLE/GLES: 4K sharp-bilinear presents in 0.14 ms mean / 0.47 ms p99 with 0 renderer allocations/frame on the RTX 4070; a 320x240 Accurate capture matches the CPU oracle on both backends.
  • Optional Vulkan: FiveDo.Video.Vulkan is an offscreen Vulkan 1.0 two-stage presenter with build-time SPIR-V shaders, sampled R16_UINT/CLUT/parameter textures, accurate/nearest/sharp-bilinear/CRT/xBR/FSR/HQx filters, GPU timestamp queries, and a host readback buffer. xBR-lv2 runs as a 4x GPU pass; FSR 1.0 runs EASU plus RCAS; HQx 2x/3x/4x run as a GPU rule-table pass over the palette texture. --vk-probe writes the shared GPU probe JSON and PNG captures; --vk-window-probe creates a native Win32 swapchain, presents through FIFO vsync, and writes the same report with offscreen: false; --gl-backend vulkan hosts the optional desktop Win32 child surface. The 320x240 Accurate probe matches the CPU oracle (0xD6BD89C5, 0 different pixels); 4K sharp-bilinear measures 0.24 ms mean / 1.06 ms p99 present and 0.075/0.091 ms GPU, 4K xBR 0.32/0.80 ms present and 0.13/0.14 ms GPU, 4K FSR 0.60/1.30 ms present and 0.40/0.42 ms GPU, and 4K HQx 2x/3x/4x 0.26/0.41, 0.29/0.34, and 0.34/0.52 ms present with 0.10/0.12/0.15 ms GPU, all with 0 managed bytes/frame on the RTX 4070. The 4K window probe reports 0.077/0.088 ms GPU and 5.96/6.53 ms present (including the 165 Hz FIFO wait), 0 managed bytes/frame. Compositor-embedded/X11/Wayland/Metal surfaces are future work. See Vulkan.md.
  • Renderer robustness: GL/GLES capability parsing reports the API and version, fullscreen/F11/Escape work, GL readback screenshots are implemented, and GL init failure or context loss switches the shell to the CPU WriteableBitmap presenter. An optional FPS/status overlay is available in the settings dialog.
  • Desktop shell: the game library scans configured folders through the 4DO database, previews sibling cover art, and opens the selected disc. Library favorites, play time, launch history, and an optional offline metadata overlay persist in library-state.json/library-metadata.json; the game list adds star toggles, metadata/play-time text, collection filters, and sort orders. The settings dialog edits video, audio, emulation, and paths; keyboard bindings for all eight players can be captured with automatic conflict removal and reset-to-defaults; gamepads are mapped through a Gamepad tab (Windows XInput, Linux evdev). Save states use ten numbered slots with next/previous, a list/save/load/delete window, JSON metadata, and PNG thumbnails; screenshots are PNG files named by game, time, and frame. File > Disc Browser... lists the current disc's Opera tree and extracts a file, directory, or the whole disc; File > Disc Info... shows the volume, game ID, database metadata, and automatic fixes. File > BIOS / NVRAM Manager... lists the BIOS library, sets BIOS1/BIOS2 and per-game assignments, and backs up, imports, resets, or deletes shared/per-game NVRAM images. The shell has a first-run guide, error dialogs, and an in-app diagnostic log viewer with level/category/search filters. The app accepts a positional disc path or --disc <path> and ships a Linux .desktop entry; Windows/macOS associations remain installer work. Per-game 4DO fix flags are applied automatically when a disc opens; explicit masks can be set in the settings dialog. The shell dispatches the configurable InputProfile.Hotkeys bindings (F5 save, F8 load, F6/F7 previous/next slot, F9 frame advance, F10 reset, F11 fullscreen, F12 screenshot, Escape pause by default) before forwarding keys to the emulated pad. Emulation > Record Input captures player-1 input to a replayable script. Loading a state whose recorded disc hash differs from the current disc reports a warning in the status line and log instead of loading silently.
  • Logging and bug reports: Serilog is configured once by each host through FiveDo.Diagnostics.Logging. The pipeline writes rolling daily files, stderr (desktop), the in-memory Log Viewer, and an asynchronous bug-report sink for Warning-or-higher events. Reports include the required environment, error, and exception blocks; submission is deduplicated for five minutes, capped at ten per minute, and disabled in CI/test runs. FIVEDO_BUG_REPORTS=0 disables it and FIVEDO_LOG_DIR redirects the files.
  • Cheats: per-game JSON cheat files are applied before every frame by FiveDoMachine.BeforeFrame, so state loads, rewind replay, and fast-forward keep the configured entries active. The engine supports raw byte writes, typed write8/write16/write32 writes, memory conditions, player-1 button conditions, and and/or/not combinations. Emulation > Cheats... and Settings > Paths > Cheats manage files; the CLI exposes cheat list|add|enable|disable|remove|import|export and run --cheat/--cheat-code/--cheat-auto. See Cheats.md.
  • Per-game fixes: GameCompatibility maps the 4DO game IDs to FixMode bits. The CLI exposes --fix-mode 0xMASK and --no-game-fixes, and run --json / info report the applied mask. Alone in the Dark's FIX_BIT_TIMING_6 is applied automatically; the current 3,000-frame boot keys are in TestCorpus.md 1.1 after the audio-clock rebaseline.
  • Tooling: CLI run, trace, bench, selftest, info, disc dump/tracks (disc tracks covers ISO/BIN/CUE as well as CHD), fs list/extract (the first-party Opera parser covers ISO/BIN/CUE/CHD), disasm, plus KPRINT, CD, CLIO, FIQ, RAM-write watchpoints, RAM/DSP dumps, DSP-instruction rings, CEL command/geometry traces, PBUS packet traces, CPU instruction-ring diagnostics, scripted input, NVRAM load/save, periodic frame capture (--capture-dir/--capture-every), RAM patches at run start (--poke ADDR=HEX), an experimental Road Rash road-following driver (--autopilot road), an opt-in bounded disc prefetch cache (--prefetch), and RAM dumps at trace stops. run --json reports the source size, VDL modulo, road coverage, autopilot state (including autopilotMenuAdvance), and CD status/poll for mode/crop, driver, and wait-state investigations.
  • Audio: the DSP executes in every run, including headless runs with no host audio sink. The AudioFolio semaphore handshake works, instrument programs run, and the corpus produces non-silent 44.1 kHz stereo WAV output. The DSP sample clock is now exact: 44,100 samples per emulated second (600 frames = 441,015-441,020 samples, 10.000 s), replacing FreeDO's 8.24 accumulator that ran ~0.7% fast (444,097 samples / 10.070 s). Road Rash's 6,000-frame capture has the same peak, RMS, and spectral peaks as before; only the timebase changed. Live output uses a bundled OpenAL Soft 1.23.1 runtime (no system OpenAL install required), makes the context current on the audio thread, pads underruns, and reports dropped/underrun/queued counters. The fill loop now rents only buffers OpenAL has released (the old queued-count index rewrote buffers still attached to the playing source and caused constant crackle), and FIVEDO_AUDIO_TAP=<prefix> records the exact queued PCM plus a per-buffer timing log. CoreRunner paces frames against the audio clock (with a 60 Hz fallback), the CLI can play live with --audio openal [--realtime], and audio analyze <wav> --json reports peak/RMS/clipping/DC statistics. Exact-clock 600-frame captures: Wolfenstein 18,100 / 2,168, Lemmings 8,192 / 1,661, Gex 12,964 / 2,089, Road Rash 32,768 / 11,012 (2,036 clipped), Theme Park silent; see TestCorpus.md 1.5a2. The trimmed win-x64 and linux-x64 CLI and the opt-in NativeAOT CLI open the same backend.
  • Input: joypad, flightstick, mouse, lightgun, arcade lightgun, and Orbatak trackball devices are supported on six PBUS ports. The legacy 4DO packet remains the default for all-joypad configurations; selecting a peripheral emits Opera's daisy-chain records. Settings > Emulation > Input devices and the CLI --input-device select devices; --pbus-format opera forces the chain layout. The desktop pointer drives the first mouse/lightgun/trackball port and shows a crosshair for lightguns. See Architecture.md §9.
  • Input scripting: the CLI accepts --input-script <path> with frame button[,button...] lines for repeatable controller taps. Input has been validated in Road Rash (race state/wreck), Wolfenstein (menus, gameplay, level start), Alone (scene/menu/intro), Lemmings (menu and passcode entry), and Space Ace (X menu, A play).
  • Tests: 1381 passing (dotnet test -c Release), including high-resolution CEL and VDLP regression coverage, RAM-patch and Road Rash autopilot contracts, CLI process tests, zero-allocation uploader/mailbox assertions, OpenGL capability parsing, bounded prefetch-cache tests, game-fix mapping, save-slot metadata/roundtrips and thumbnails, keyboard/gamepad binding capture, gamepad mapping, log filtering, source-generated CLI JSON contracts, frontend startup-argument parsing, renderer backend selection, GPU-probe capture/parity contracts, shader source invariants, Vulkan xBR/FSR/HQx split-frame orientation, save-state disc-hash warnings, hotkey routing, input-recorder coalescing/roundtrip, ISO/BIN disc tracks, Opera FS parsing (multi-block directories, burst/gap reads, invalid headers, and CLI ISO extraction), and BIOS catalog/scan/selection/override tests.
  • Packaging: CI builds/tests/formats and publishes one bundle per system (5do-win-x64.zip, 5do-linux-x64.tar.gz, 5do-osx-x64.tar.gz). Each bundle contains the CLI and desktop app as framework-dependent single-file apphosts, the README/license/notices/whatsnew, and a generated PACKAGE.txt. The .NET 10 runtime is required on the host and is not bundled; PDB and XML documentation files are removed. Native OpenAL Soft/Skia/ANGLE libraries stay beside the apphosts for LGPL replacement. The CLI also has an opt-in NativeAOT win-x64 publish (-p:FiveDoNativeAot=true) that is CI-smoke-tested with --version, selftest, bench, info, fs list, a golden 3,000-frame boot, and live OpenAL audio. The frontend has an opt-in NativeAOT single-file publish that passes --smoke-test and the WGL/ANGLE parity probes. See packaging/README.md.

1a. Latest compatibility sweep (2026-10-30)

The first full loadability pass over both collections is complete. It booted every disc for 3,000 frames and recorded JSON/PNG evidence under artifacts/compat-full/. The committed summary is GameCompatibilityResults.md.

Result Count Meaning
Loadable/Rendered 593 Completed 3,000 frames (18 rows retried at 20,000-65,000 frames); final frame is non-black.
Loadable/No video 1 Completed 65,000 frames; still stuck on a boot screen.
Not Loadable 0 No CLI error or timeout remains.
Total 594 589 main collection + 5 MAME software-list images.

The 2026-11-01 extended-boot retry reran every one of the 18 non-rendered rows. The five former Not Loadable rows (Crime Patrol, Mad Dog II, Mirai Shounen Conan, Orbatak, and Shootout at Old Tucson RE3) now complete and render at 20,000 frames with panafz10, and 11 of the 13 former black/blank rows render at 20,000 frames or at 65,000 with the Anvil BIOS. The two remaining rows are known Anvil boot-screen hangs: Crash 'n Burn (Japan). WorldCup Super Stadium now renders when the FZ-10J BIOS is paired with its Kanji font ROM (BIOS2).

The 2026-11-01 retry reran all 13 former Loadable/No video titles. Ishida Yoshio Kudan no Igo Seiha, both Konpeki no Kantai discs, all three Myst discs, Olympic Summer Games, and all three Slayer discs render at 20,000 frames with panafz10; Powers Kingdom (Europe) renders its live-action intro at 65,000 frames with panafz10e-anvil. Only Crash 'n Burn (Japan) remains on the Anvil boot screen at 65,000 frames; WorldCup Super Stadium renders with the FZ-10J BIOS and its Kanji font ROM.

Loadable/Rendered is not a final compatibility grade; it still needs the visual and input review described in CompatibilitySweep.md.

2. Real-game status (local corpus, panafz1.bin)

The table below is the curated five-title showcase set. The full local collection (594 images across J:\Panasonic 3DO and G:\MAME\MAME Software List CHDs\3do) is being swept in M13; the working matrix is GameCompatibilityList.md.

The 2026-10-24 comparison used the same BIOS and an ISO extracted from the same CHD on both sides. Opera is the bundled Windows OperaCore build in References/CPP_SimpleGui; it is a behavioral reference, not product code.

Title Our C# port Opera reference
Lemmings (USA) Boots, renders, plays DSP audio; reaches the title screen and the passcode screen; with --high-res the 640x480 title renders all four buttons and the full logo, matching the reference. The Fun 4 code HNLHCIOECW is now typed entirely with the D-pad (docs/scripts/lem-fun4-dpad.txt); gameplay starts at 23,000 frames (PC 0x00083FE8, hash 0x63C02DED) and roundtrips (PC 0x000230A4, 0x4CA339CD twice). 600-frame WAV evidence in TestCorpus.md 1.5a Not re-tested this pass; expected to pass
Space Ace (USA) Boots, renders, plays DSP audio; ReadySoft player runs the animated demo past 12,000 frames. The interactive path is now validated: X returns to the title/menu, A starts the scene; the 15,000-frame state roundtrips (0x7E780D93 -> PC 0x000284EC, 0x70B16422 twice) and the gameplay WAV is non-silent Plays the demo intro for 60 s, no runaway
Road Rash (USA) Boots, renders, plays DSP audio; the 3D race renders correctly and the attract race runs past 60,000 frames with no allocator error; B accelerates, A brakes, the D-pad steers, C attacks, and P pauses. The CD media-access pulse fixes the old ~42k/~92k READ_ID-only wait: those screens advance to BUSTED, the race-select menu, and a new race. The autopilot now passes GAME LOADING/Get ready and cycles race -> wreck -> menu -> new race to 100k, but still does not finish a full 5.5-mile race Renders the intro FMV, then a race
Wolfenstein 3D (USA) Boots, renders, plays DSP audio; A opens the preview menu, selecting a level starts a fresh game (FLOOR 1-1), the player can walk/turn/fire, and save/load roundtrips in-game (0xF568EBCA -> PC 0x000782B8 twice); 600-frame WAV evidence captured. The fresh-game load advances with A pulses (character select -> FLOOR 1-1 at 64,000, hash 0x2F5B9864) Renders the intro for 60 s
Alone in the Dark (USA) Boots, renders, plays DSP audio; reaches the in-game staircase/room with an upright character; the Up/Left/A/Down sequence reaches the Begin/Resume menu, A starts the intro scene, A pulses pass the intro, and a scripted room movement sequence roundtrips (0x44C7576B -> two 120-frame continuations at PC 0x0008BF00, 0xE490C8C7) Renders the intro/loading for 60 s

Additional corpus (2026-10-26). Seven more titles were booted, rendered, and save-state-roundtripped with panafz1.bin: Gex, Need for Speed, Super Street Fighter II Turbo, Total Eclipse, Killing Time, Theme Park, and Return Fire. Gex, Total Eclipse, and Super Street Fighter II Turbo show gameplay/attract scenes; Need for Speed, Killing Time, Theme Park, and Return Fire render menus. Six produce non-silent 44.1 kHz stereo audio; Theme Park's captured menu is silent. Hashes and commands are in TestCorpus.md 1.7 and the matrix in Compatibility.md §10.

The CD disc-end fix on 2026-10-23 is what unblocked Road Rash, Space Ace, and Wolfenstein: InitCd no longer double-adds the 150-frame lead-in, so the CD DIPIR accepts discs whose volume block count is less than the physical track length.

The 2026-10-25 eighth pass fixed a real-content renderer bug: the GPU palette shader used the low five bits of each pixel as a single CLUT index for all three channels, so games appeared grayscale. It now uses the red, green, and blue indices independently and matches the CPU frame assembler. Lemmings gameplay was captured through sharp-bilinear and xBR with correct colors. The 2026-10-29 pass removed the former CPU-expanded xBR source; xBR now samples the GPU palette texture directly.

The 2026-10-24 session found and fixed the common blockers:

  1. DSP execution. AudioSink?.PushSample(Dsp.Loop()) skipped the DSP entirely when the host had no audio sink, which is the default headless path. The machine now always executes the DSP; the sink only receives samples.
  2. DSP immediate operands. The loader shifted un-justified immediates left by two instead of zero. The AudioFolio tail program's Sema4Status & 8 test became Sema4Status & 0x20, so the DSP always skipped the ARM command. The fix uses the reference's IMMEDIATE << (JUSTIFY & 3). SAudio -6996 is gone and the DSP now runs instrument code.
  3. DSP multiplier rounding. MUXB=3 now masks the product ((MULT1 * (Y >> 15)) & ~1) instead of masking the shifted accumulator first.
  4. CCB field order. The parser now reads position, size, perspective, and pixel-processor control before PRE0/PRE1, matching FreeDO. The old order mis-decoded every 3D CEL; Road Rash's race produced a bogus 1987x583 CEL and appeared to hang. The rasterizer's missing cnt_cross > 2 x-point sort was also ported.
  5. NVRAM initialization. The CLI now loads a fresh 32 KB NVRAM volume by default (and --nvram <path> persists it). Without it, Alone in the Dark saw zero free save slots, displayed its save-space warning, and shut down. It now reaches the in-game room/inventory scene.
  6. Arbitrary-CEL transposition. TexelDrawArbitrary passed (i, j) to (x, y) read/write helpers, so every asymmetric quad was written transposed. Road Rash's race CELs overwrote the OS VRAM free list at 0x2FC000 (AllocMem error, bad node address) and the race froze; Alone's character was drawn horizontally. Passing (j, i) fixed both. The VDLP bitmap copy now uses the modulo-derived line width (no 320-pixel crop), and state loads restore the VDL pixel buffers.

Space Ace's former frame-681 crash was a downstream consequence of the failed AllocInstrument: the player copied a decoded buffer to a null pointer, wiping the ARM vectors. It now runs past 12,000 frames.

Gameplay is now reachable and renderer-correct. Road Rash's 3D race renders like the Opera reference and runs past 60,000 frames; Wolfenstein 3D starts a fresh game, walks/turns/fires, and its save states roundtrip; Alone in the Dark's in-game character is upright and its Begin/Resume menu is reachable; Lemmings completes a Fun 4 passcode (HNLHCIOECW) end-to-end into gameplay with save/load roundtrip, and its 640x480 high-resolution title matches the reference; Space Ace starts its interactive scene (X -> menu, A -> play) and plays. Per-title save-state, PNG, and WAV evidence is recorded in TestCorpus.md 1.5a.

The ninth pass corrected the Road Rash diagnosis and fixed the wait: the ~42k and ~92k stops were a static READ_ID-only CD wait, not wall pins or driving failures. The CD device now raises a media-access event (0x40) after two consecutive READ_ID polls with an idle gap; this advances those screens to BUSTED, race select, and a new race. run --json exposes cdStatus/cdPoll, and the autopilot reports autopilotMenuAdvance while it pulses A through the post-race flow. The 3,000-frame keys were later re-baselined by the exact audio clock; see TestCorpus.md 1.1. From the 60k state the autopilot now starts a new race at 80k and cycles race -> wreck -> menu -> new race to 100k; the remaining Road Rash work is finishing a full race without wrecking, not the CD wait. That is post-1.0 backlog; it is not a v1 release gate.

3. Performance snapshot

Release build, development machine, 600 frames, CPU side only:

Case Total p50 p95 p99 Allocations
Synthetic 1.250 ms/frame 1.218 ms 1.434 ms 1.638 ms 0 bytes/frame
Lemmings CHD 1.281 ms/frame 1.088 ms 2.005 ms 3.398 ms ~5.9 KB/frame (CHD decode)
Lemmings CHD, high-res 1.605 ms/frame 1.398 ms 2.497 ms 4.598 ms ~5.8 KB/frame (CHD decode)
Road Rash 3D race (100k state) 4.20 ms/frame 2.15 ms 7.13 ms 7.40 ms 0 bytes/frame

This leaves ~15.4 ms of headroom against the 16.67 ms frame budget for the synthetic case. The 2026-10-31 hot-path pass reduced the Road Rash race from 5.46 to 4.20 ms mean and from 13.14 to 7.40 ms p99 (frame hash unchanged). The core projects are also compiled with JIT optimizations in Debug, so an IDE Debug run is no longer limited to ~40 fps. CI now runs the synthetic benchmark with --gate: it fails on any steady-state allocation, core p99 > 8 ms, total mean > 8 ms, or total p99 > 16.667 ms, and uploads the JSON result as the baseline artifact. The benchmark's --baseline option also implements the same-machine >10% relative comparison (core p99, total mean, total p99; exit code 4); wiring it to CI awaits a stable/self-hosted runner. Tier B desktop numbers and GLES timing are not measured yet. CHD decode allocations are host-side and are attributed to the core stage; the synthetic core/pack/mailbox/read path is allocation-free in steady state.

The desktop renderer uses a GL_TIME_ELAPSED query and reports GPU/present time in the FPS overlay. The fourteenth pass added a true-4K offscreen --gl-probe and measured the presentation path on Windows 11 (Ryzen 9 7900, RTX 4070, OpenGL 4.0):

4K case (3840x2160) Present mean Present p99 GPU mean GPU p99 Renderer alloc
Sharp-bilinear 0.080 ms 0.208 ms 0.086 ms 0.089 ms 0 B/frame
Accurate 0.082 ms 0.178 ms 0.091 ms 0.287 ms 0 B/frame
CRT 0.147 ms 1.157 ms 0.191 ms 0.193 ms 0 B/frame
xBR 4x 0.283 ms 1.466 ms 0.185 ms 0.324 ms 0 B/frame
FSR (EASU + RCAS) 0.132 ms 1.191 ms 0.441 ms 0.483 ms 0 B/frame

The xBR row now uses the GPU-expanded palette texture directly; the previous path uploaded a CPU FrameAssembler copy. The optional Vulkan offscreen backend was also measured on the same machine: 4K sharp-bilinear present 0.24 ms mean / 1.06 ms p99, GPU 0.075/0.091 ms, 0 managed bytes/frame; 4K xBR present 0.32/0.80 ms, GPU 0.13/0.14 ms; 4K FSR present 0.60/1.30 ms, GPU 0.40/0.42 ms; 4K HQx 2x/3x/4x present 0.32/0.83, 0.32/0.48, and 0.40/0.69 ms with GPU 0.10/0.11, 0.12/0.13, and 0.15/0.16 ms; 1080p sharp-bilinear present 0.17/0.79 ms, 1080p xBR 0.21/0.38 ms, and 1080p FSR 0.27/0.78 ms, all with 0 managed bytes/frame. GPU time excludes the optional per-frame readback. Its 320x240 Accurate output matches the CPU oracle (0xD6BD89C5, 0 different pixels). See Vulkan.md.

The platform swap interval is wired through SwapInterval.TryCreate (WGL/GLX/EGL); the Windows WGL path resolves WGL_EXT_swap_control and reports vsyncActive: true when requested. The Avalonia compositor paces the window independently, so the probe reports ~64 fps throughput even when vsync is requested; the renderer cost itself is far below the 16.67 ms budget. DisplayRefreshRate (Settings > Video, default 60) re-presents the latest frame above 60 Hz when the user wants a faster swap rate; it does not add motion because the emulated 3DO still outputs 60 fields per second.

4K60 feasibility

  • Desktop (Tier A): measured. The core is ~1.25-1.6 ms/frame; the 4K presentation pass is one fullscreen textured quad after native-resolution palette expansion. The RTX 4070 measured GPU p99 0.089 ms and present p99 0.208 ms for sharp-bilinear, leaving ample headroom for 60 Hz. The remaining risk is the Avalonia/WGL compositor and frame pacing, not the emulator core.
  • Tier B desktops: unmeasured. The presenter should handle 1080p60; integrated GPUs and older CPUs may be the limiting factor in complex scenes.
  • 4K internal CEL rendering is not a v1 goal. The target is 4K output, not 4K internal rendering.

CPU p99 and per-stage allocations are measured, and Tier A 4K GPU/present p99 is now recorded. Remaining performance work is post-1.0: GLES 3.1 and DPI validation, Tier B timing, and the CI relative gate. WebAssembly, Android/iOS, and Raspberry Pi are out of scope. See Performance.md.

4. How far from a release?

The original 157 developer-day plan assumed full compatibility hardening and a full platform matrix. v1.0 adopts a documented-compatibility release scope instead. Ship the platforms and titles that are validated, publish the per-title matrix and known issues, and move the rest to the post-1.0 backlog. Most emulators release this way; full compatibility is not a v1 requirement.

As of 2026-11-02 the v1.0.0 documented-compatibility release is published on GitHub. M13 loadability is complete (594 tested, 593 render, 1 known boot failure, 0 not loadable), legal/attribution sign-off is recorded, and M14 is post-1.0 research. Linux x64 CLI/frontend smoke and the full test suite passed (see Compatibility.md §10.3); the GL probe timed out under llvmpipe, so Linux GPU presentation is preview/best-effort. The release artifacts and checksums are attached to the GitHub release.

Release workstream State Remaining
M0-M8 core, CPU, boot, MADAM, DSP, CD, input, save states Functionally complete 0
M9 Avalonia frontend Feature-complete for the tested workflow 0 (optional polish is post-1.0)
M9R OpenGL presenter Windows WGL + ANGLE validated; CPU fallback, filters, parity, 4K probe 0 on Windows; Linux x64 full suite passed, GL probe timed out under llvmpipe
M9P performance/pacing Tier A measured, zero-allocation steady state, CI gate 0 (Tier B/Pi and hot-path are post-1.0)
M10 packaging Unified Windows/Linux/macOS x64 bundles, framework-dependent single-file apps, optional NativeAOT 0 (installers post-1.0)
M11 compatibility 12 titles boot/render; five have scripted gameplay evidence; matrix and known issues documented 0
M13 compatibility sweep Loadability complete: 594 tested, 593 render, 1 known boot failure M13.5/M13.6 post-1.0; Linux x64 full suite passed
M14 high-res CEL research Research plan written; post-1.0 Not a v1.0 gate
M11 release/legal OpenAL notice/source offer/relinking record and license text ship in every bundle; FreeDO modified GNU LGPL, donationware Legal sign-off recorded 2026-11-02

Current estimate: roughly 100% complete against the documented-compatibility release scope. v1.0.0 is published on GitHub with the release artifacts.

Release decisions (2026-11-02): Windows x64 is validated; Linux x64 CLI/frontend smoke and the full test suite passed (GPU presentation is preview/best-effort); the current release version is 1.6.0; installers, signing, and notarization are post-1.0; the compatibility set is the 12 documented titles, with five showcase titles carrying scripted gameplay evidence. The distribution model is recorded as donationware (non-commercial, source-available).

The project license is recorded in LICENSE as the FreeDO modified GNU LGPL, so release builds are donationware: non-commercial, source-available, and funded by voluntary donations rather than sales. The project-owner legal/attribution sign-off is recorded in ReleaseLegalReview.md. The OpenAL Soft LGPL-2.0-or-later binary-distribution obligation is handled: the notice, complete license text, three-year source offer, and relinking instructions ship in OpenALSoft.md and OpenAL-Soft-COPYING.txt. See Licensing.md, ReleaseLegalReview.md, and THIRD-PARTY-NOTICES.md.

Gameplay-depth and full-compatibility work is post-release, not a current gate: finishing Road Rash's 5.5-mile race, scripting Alone's inventory, adding gameplay scripts for the other seven titles, and further CD/CLIO/MADAM/ARM timing triage. M13.5/M13.6 visual/high-resolution review and M14 research are post-1.0. The current v1.0 gate is release mechanics (version bump, tag, and artifacts).

5. Current path (compatibility and video quality)

P0 - v1.0 release work

  1. M13 loadability sweep: complete. 594 discs tested: 593 Loadable/Rendered, 1 known Loadable/No video failure (Crash 'n Burn Japan), 0 Not Loadable. See GameCompatibilityResults.md.
  2. Failure triage: complete. The five former not-loadable rows and 12 black/blank titles are resolved or explained; Crash 'n Burn (Japan) remains the only known boot failure. WorldCup Super Stadium renders with the FZ-10J BIOS plus its Kanji font ROM.
  3. Linux x64 validation: complete (2026-11-02). The full test suite (1,473/1,473), CLI selftest, bench, info, bios info, a 3,000-frame Wolfenstein boot, non-silent audio, and the frontend --smoke-test passed on Linux Mint 22.3. The GL probe timed out under llvmpipe, so Linux GPU presentation is preview/best-effort. See Compatibility.md §10.3.
  4. Release v1.0.0: complete (2026-11-02). The version is bumped, the source is tagged v1.0.0, the unified Windows/Linux/macOS x64 bundles and checksums are attached to the GitHub release, and the release-owner/QA sign-off is recorded.
  5. M13.5/M13.6 visual and high-resolution review (post-1.0). Inspect every rendered final frame, record Unplayable/Playable/Good, and run the -HighRes comparison.

P1 - research and later work

  1. M14 high-resolution CEL research (post-1.0): measure the current mode, prototype true 640x480 and 2x supersampling, and write the GPU CEL feasibility note.
  2. Compatibility fixes: only after the failure classes are triaged and compared with the 4DO/Opera reference.
  3. Gameplay depth (deferred): finish a full Road Rash race, script Alone's inventory, and add input scripts for the other seven titles.
  4. Platforms/performance/packaging (post-1.0): Pi/GLES 3.1, arm64 hardware, DPI, Tier B timing, installers/signing, file associations, and the optional M12 JIT.
  5. Release status: v1.0.0 is published on GitHub with the artifacts attached. Legal/attribution sign-off is recorded, CI is green, and Linux x64 full suite passed. Post-1.0 work remains.

6. Next session

v1.0.0 is published on GitHub (source tag v1.0.0, unified Windows/Linux/macOS x64 bundles and checksums attached). Linux x64 validation passed on Linux Mint 22.3: the full test suite (1,473/1,473), CLI selftest/bench/info, a 3,000-frame Wolfenstein boot, non-silent audio, and the frontend --smoke-test. The GL probe timed out under llvmpipe and is a documented preview/best-effort limitation. The BIOS-aware retry is complete: 593 of 594 discs render, and only Crash 'n Burn (Japan) remains stuck on the Anvil boot screen. The next session is post-1.0:

  1. The v1.0.0 GitHub release is published; collect feedback and compatibility reports.
  2. Run the M13.5/M13.6 in-game/visual and high-resolution review.
  3. Start M14 from HighResolutionCelResearch.md.

Do not re-enable the null audio-sink short-circuit, do not change the XBUS DMA byte order (the raw IRam order is correct), do not port opera_cdrom.c wholesale, and do not redo the trimmed CLI/frontend or source-generated JSON work; see Trimming.md. Do not treat full-race completion, scripts for the other seven titles, Pi/arm64 validation, or installers as current gates; they are later backlog.

7. Reproduce

dotnet build -c Release
dotnet test tests/FiveDo.Tests -c Release

# The corpus roots are machine-specific; set them once per shell.
$env:FIVEDO_GAMES = "J:\Panasonic 3DO"
$env:FIVEDO_MAME  = "G:\MAME\MAME Software List CHDs\3do"
$env:FIVEDO_DISC  = "$env:FIVEDO_GAMES\Road Rash (USA).chd"

# Full-collection loadability sweep (3000 frames, resume-safe)
pwsh -File tools/compat-sweep.ps1 -Frames 3000 -Resume -OutDir artifacts/compat

# Regenerate the 4DO reference table after changing the collection
python tools/generate-compat-list.py

# Build/test/format
dotnet build -c Release
dotnet test tests/FiveDo.Tests -c Release
dotnet format --verify-no-changes

# Synthetic CPU-side baseline
dotnet run -c Release --project benchmarks/FiveDo.Benchmarks -- --frames 600

# Real disc baseline
dotnet run -c Release --project benchmarks/FiveDo.Benchmarks -- `
  --frames 600 --bios bios/panafz1.bin `
  --disc "$env:FIVEDO_GAMES\Lemmings (USA).chd" --label lemmings

# Boot a disc and dump the last frame
dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Lemmings (USA).chd" --bios bios/panafz1.bin `
  --frames 3000 --nvram artifacts/lemmings.ram --frame-out artifacts/lemmings.png

# Scripted input, frame sequence, and PBUS packet trace
dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Wolfenstein 3D (USA).chd" --bios bios/panafz1.bin `
  --frames 12000 --input-script artifacts/input.txt `
  --capture-dir artifacts/w3d-frames --capture-every 1000 `
  --frame-out artifacts/w3d.png --state-out artifacts/w3d.state

# RAM dump at a KPRINT stop
dotnet run -c Release --project src/FiveDo.Cli -- trace `
  "$env:FIVEDO_DISC" --bios bios/panafz1.bin `
  --frames 16000 --trace-frames 2200 --stop-kprint "AllocMem error" --ring 40000 `
  --out artifacts/rr-trace.txt --dump-ram artifacts/rr-stop-ram.bin `
  --dump-ram-offset 0 --dump-ram-length 0x300000

# Opera reference: extract the data track to ISO, then run the bundled SimpleGui
dotnet run -c Release --project src/FiveDo.Cli -- disc dump `
  "$env:FIVEDO_DISC" --sector 0 --count 328130 `
  --out "$env:TEMP\roadrash.iso"
& "References\CPP_SimpleGui\cmake-build-release\SimpleGui\SimpleGui.exe" `
  --debug-messages -n "$env:TEMP\roadrash.iso"

The desktop frontend is dotnet run -c Release --project src/FiveDo.Frontend.Avalonia. See ReferenceStrategy.md for the full differential procedure.